Study On Applications Of Neural Networks For Speech Signal Detection

In many cases of speech signal processing, the problem of speech detection is involved, such as objective assessment of speech quality in communication systems, encoding and decoding of speech signals, speech recognition and single channel speech toning etc. Actually, the practicality of speech signal processing techniques is strongly demanded in many fields, especially the objective assessment technique of speech quality to assess the performance of speech system expediently, rapidly and reliably, has become the goal of researchers world wide. As the result, the first and one of the most important tasks is to detect the jumping-off point of speech signals reliably under noisy environments.The traditional speech detection techniques are not only limited in precision, but also mostly can be efficient only under undisturbed environment like the lab or if the background noise is very limited. According to this, under the background of a cooperative plan about the objective assessment of speech quality, the study and discussion of reliable speech detection algorithms under very noisy environments, implementation and simulation analysis, are developed. Two new speech detection methods based on neural networks using Barker coded synchronization signal are presented. Simulation results show that they can work well under very noisy environments, and the precision arrives at 10 samples.The main research works carried out in this thesis are listed below.First, the traditional speech detection method based on short-time energy is discussed, including its principle and implementation. Then it is used for the jumping-off point detection of speech signals transmitted by AWGN channel. Simulation results are provided.It is conceived to introduce Barker Codes as synchronization preamble and add synchronization signal in front of speech signal to implement speech detection. The principle of this idea is presented.A new speech detection method using the MLP neural network is studied, including the structure of MLP NN and BP learning algorithm. It is then used for the jumping-off point detection of speech signals transmitted under different channel conditions. Simulation results show that it works well even if the channelcondition is very bad.Another new speech detection method using the AD ALINE neural network is discussed, including its structure, learning algorithm and simulation results for the detection of speech signals transmitted under different channel conditions.